three-tier-memory_skill

Overview

This skill implements a Three-Tier Memory system for AI agents: short-term sliding window, medium-term automatic summaries (LLM), and long-term semantic vector retrieval. It helps manage conversation history, create automatic summaries, and build a searchable personal knowledge base. The design balances real-time context with compressed summaries and scalable semantic retrieval.

How this skill works

Short-term memory holds the most recent N messages in a FIFO sliding window for immediate context. When token or window thresholds are reached, medium-term memory automatically generates LLM-based summaries saved as persistent notes. Long-term memory indexes summaries and important entries into a vector store for semantic retrieval and RAG-style augmentation. Hooks can auto-save sessions and trigger summaries, and a CLI provides init/add/search/status operations.

When to use it

• Maintaining relevant short-term conversation context during multi-turn dialogues.
• Automatically compressing long chat histories into concise, retrievable summaries.
• Building a personal knowledge base that supports semantic search and retrieval.
• Augmenting agent responses with relevant long-term facts via RAG.
• Backing up and archiving conversational data with layered retention policies.

Best practices

• Keep short-term window small (default 10 items) to reduce prompt size and cost.
• Tune medium-term summary thresholds to balance detail and storage.
• Use the vector store for semantic queries and set top_k to a sensible default (e.g., 3).
• Enable session hooks to automatically persist end-of-session context and reduce manual steps.
• Protect and encrypt sensitive long-term entries; avoid storing raw secrets in long-term memory.

Example use cases

• Customer support agent that keeps recent conversation context, summarizes long interactions, and retrieves archived solutions.
• Personal assistant that remembers user preferences short-term, summarizes meeting notes, and finds related past notes via semantic search.
• Research assistant that archives ideas and drafts into a vector store for later retrieval and citation.
• Chatbot integrated with OpenClaw hooks to auto-save sessions and trigger summaries when windows fill.
• Backup pipeline that archives all versions of skills or conversation artifacts and supports fast semantic lookup.